Pregnancy loss or fetal death is defined as the death of the fetus before complete expulsion from its mother( Reference Sims and Collins 1 ) and is subdivided into spontaneous abortion (or miscarriage) and stillbirth. The estimated global incidence of spontaneous abortion was 12 to 15 % among clinical pregnancies( Reference Garcı́a-Enguı́danos, Calle and Valero 2 ), while the stillbirth rate has been estimated to be 19 per 1000 births( Reference Lawn, Blencowe and Pattinson 3 ).
Caffeine is the most commonly used psychoactive substance, found mainly in coffee, tea, cola soft drinks and cocoa( Reference Maslova, Bhattacharya and Lin 4 , Reference Eteng, Eyong and Akpanyung 5 ). Caffeine is absorbed rapidly upon ingestion and readily passes the placental barrier( 6 , Reference Arnaud, Bracco and Sauvageat 7 ). Accumulation of caffeine metabolites in the fetal brain has been documented, probably due to the absence of the main caffeine-metabolizing enzyme, cytochrome P450 1A2 (CYP1A2), in both the placenta and the fetus( 6 , Reference Aldridge, Aranda and Neims 8 , Reference Sengpiel, Elind and Bacelis 9 ). The rate of caffeine metabolism decreases from the first to third trimester and the half-life of caffeine doubles in the mother during pregnancy( Reference Knutti, Rothweiler and Schlatter 10 , Reference Brazier, Ritter and Berland 11 ), leading to higher exposure for the fetus to maternally ingested caffeine. Exposure to caffeine may lead to vasoconstriction in the uteroplacental circulation, which can in turn affect fetal growth and development( Reference Bech 12 , Reference Kirkinen, Jouppila and Koivula 13 ).
Spontaneous abortion and stillbirth may have different underlying physiology and risk factors. For instance, while spontaneous abortion is caused mainly by implantation defects and chromosomal, autoimmune and endocrine abnormalities, the important causes of stillbirth include placental insufficiency and abruption, and pre-eclampsia( Reference Regan and Rai 14 , Reference O’Neill, Kearney and Kenny 15 ). However, spontaneous abortion and stillbirth share some common risk factors such as advanced maternal age, smoking and BMI( Reference O’Neill, Kearney and Kenny 15 ). It is currently unclear if caffeine has a differential influence on spontaneous abortion and stillbirth through different mechanisms.
The possible harmful effects of caffeine intake on fetal and birth outcomes warrant evaluation as many women consume caffeine-containing foods and beverages during pregnancy( Reference Maslova, Bhattacharya and Lin 4 , Reference Frary, Johnson and Wang 16 , Reference Peck, Leviton and Cowan 17 ). Considerable inconsistency exists in the literature concerning the relationship of maternal caffeine intake and pregnancy loss( Reference Peck, Leviton and Cowan 17 ). While high maternal caffeine intakes were more consistently associated with a higher risk of pregnancy loss, results have been mixed for moderate and low caffeine use, probably due to difficulties in measuring caffeine intakes and differences in study settings and participants. A recent meta-analysis summarized the effect of maternal caffeine intake during pregnancy on a variety of birth outcomes( Reference Greenwood, Thatcher and Ye 18 ). However, that analysis included only a dose–response analysis and not a categorical analysis and evaluated only a limited number of potential sources of heterogeneity in study results. We therefore conducted a categorical and dose–response meta-analysis on the association of maternal caffeine intake during pregnancy and risk of pregnancy loss in prospective studies.
The present meta-analysis was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)( Reference Moher, Liberati and Tetzlaff 19 ) (see online supplementary Table S1) and Meta-analysis of Observational Studies in Epidemiology (MOOSE)( Reference Stroup, Berlin and Morton 20 ) guidelines (see online supplementary Table S2).
Two investigators searched MEDLINE and SCOPUS databases through 30 January 2015 with no language restriction. SCOPUS is an abstract and citation database of peer-reviewed literature that includes all the contents from the EMBASE database( Reference Elsevier 21 ). The search was based on combinations of synonyms for caffeine (including its chemical name, coffee and tea) and pregnancy loss (including spontaneous abortion, miscarriage, stillbirth and fetal death). The detailed search strategy is shown in online supplementary Material S1.
Studies were included if they met the following criteria: (i) the study was an original, peer-reviewed study (i.e. not review articles or meeting abstracts); (ii) the study was a prospective cohort study or nested case–control study; and (iii) the authors reported the risk estimates of pregnancy loss associated with maternal caffeine intake (estimated total caffeine intake or coffee intake as a proxy for total caffeine intake) during pregnancy or the periconceptional period. We excluded studies that presented crude estimates only or did not consider potential confounding by smoking, and studies conducted in unhealthy populations (e.g. type 1 diabetes or infertility).
Pregnancy loss included both spontaneous abortion and stillbirth. Spontaneous abortion was defined as spontaneous loss of a fetus before the 20th week of pregnancy and stillbirth was defined as the death of a fetus during birth or during the last half of pregnancy. However, the definitions for stillbirth and miscarriage are not standard across the world, with miscarriage being defined as pregnancy loss occurring at less than 20–28 weeks of gestation.
The study selection was conducted independently by two authors. We also considered non-English articles with help from colleagues who are proficient in these languages. The inter-rater agreement was fairly good (κ=0·66; κ values that range from 0·61 to 0·80 indicate ‘substantial agreement’( Reference Sim and Wright 22 )). Discrepancies were resolved by discussion with a third investigator.
For included studies, information on participants, study design, measurement of exposure and outcome, effect estimates and their se (or related statistics) were extracted independently by two investigators using a standardized extraction form. Discrepancies were resolved by discussion with a third investigator. Study quality assessment was conducted by considering characteristics such as study design, number of cases and participants, method of exposure assessment and adjustment for confounders. Klebanoff et al.( Reference Klebanoff, Levine and DerSimonian 23 ) did not report numbers for effect estimates and we estimated the effect estimates and corresponding CI from their figure.
The multivariable-adjusted OR, relative risks (RR) or hazard ratios (HR) have been used in different studies. We chose RR as a measure of risk estimates in the meta-analysis because the incident rate was low (0·4 to 25·1 %; <20 % in twelve out of fourteen included studies), and OR and HR thus approximated RR( Reference Thomas 24 ).
Different studies used different cut-off points for the caffeine intake categories. To combine the risk estimates from different categories in different studies, we assigned the median value for each category of caffeine intake. When lower and upper boundaries were presented for the category, we assigned the midpoint as an estimate of the median caffeine intake. If the upper boundary of the highest category was not provided, we assumed that the boundary had the same amplitude as the second-highest category( Reference Larsson and Orsini 25 ). If the lower boundary of the lowest category was not provided, we assumed the lower boundary to be zero( Reference Larsson and Orsini 25 ). Two studies reported coffee consumption in cups only but not total caffeine intake( Reference Bech 12 , Reference Wisborg, Kesmodel and Bech 26 ) and we estimated caffeine intake based on the commonly cited conversion method (107 mg caffeine per cup of coffee)( Reference Bunker and McWilliams 27 ). Klebanoff et al.( Reference Klebanoff, Levine and DerSimonian 23 ) reported paraxanthine concentration (a metabolite of caffeine) as the exposure and we estimated caffeine intake based on the conversion method proposed in that article. Briefly, for a 60-kg woman, caffeine intakes of 600 mg in non-smokers and 1100 mg in smokers were estimated to translate to serum paraxanthine of 1845 ng/ml. The conversion was extrapolated from the authors’ pilot data in a pregnant population with an overall lower caffeine intake( Reference Klebanoff, Levine and Dersimonian 28 ). Several studies reported results for more than one period of maternal caffeine exposure (Table 1). We used the results for the assessment period most frequently used (first trimester) in included studies. Nevertheless, we also included results from the other assessment periods in stratified analyses where possible. Although stillbirth and miscarriage can have different aetiology, as only one included study focused specifically on stillbirth and the incidence of stillbirth is low compared with miscarriage in included studies combining both outcomes, we did not separate these outcomes in our main analysis.
NA, not available; LMP, last menstrual period; RR, relative risk.
* Included for qualitative review only.
† Data used for the main analysis.
We first conducted analyses based on different levels of caffeine consumption. We identified five levels of caffeine consumption based on assigned median caffeine consumption level: (i) reference category (<50 mg/d); (ii) low caffeine consumption (50–149 mg/d); (iii) moderate caffeine consumption (150–349 mg/d); (iv) high caffeine consumption (350–699 mg/d); and (v) very high caffeine consumption (≥700 mg/d). Effect estimates of the individual studies were combined using the random-effects method as described by DerSimonian and Laird( Reference DerSimonian and Laird 29 ), which considers both within-study and between-study variations. The Cochran Q test and I 2 statistic were used to evaluate statistical heterogeneity among studies( Reference Higgins and Thompson 30 , Reference Higgins, Thompson and Deeks 31 ), and I 2 values of 25 %, 50 % and 75 % correspond to low, moderate and high degrees of heterogeneity, respectively( Reference Higgins, Thompson and Deeks 31 ).
We further conducted a dose–response analysis using the generalized least-squares trend estimation (GLST) method as described by Greenland and Longnecker( Reference Greenland and Longnecker 32 , Reference Orsini, Bellocco and Greenland 33 ), which computes the trend from the correlated log RR estimates across caffeine consumption categories. The caffeine dose used in this analysis was based on median caffeine consumption level derived using the same methods as category-based analysis described above. We performed a two-stage GLST method that first estimates study-specific slopes before deriving an overall average slope( Reference Orsini, Bellocco and Greenland 33 ), because this method allowed us to include effect estimates from studies that reported results for caffeine intake only as a continuous variable. We tested for a potential non-linear relationship between maternal caffeine intake and pregnancy loss using a restricted cubic spline random-effects model with three knots; the P value for non-linearity was obtained by testing the null hypothesis that the spline term is equal to 0.
We conducted stratified analyses and meta-regression analyses to assess potential sources of heterogeneity by different study-level characteristics. In addition, we conducted sensitivity analysis to assess the influence of each individual study by omitting one study at a time and calculating the summary RR for the remaining studies. We also excluded studies focusing on stillbirth or a combination of miscarriage and stillbirth to evaluate the influence of pooling effect estimates from different outcomes. Publication bias was evaluated with Egger’s regression test, Begg’s adjusted rank correlation test and visual inspection of the funnel plot( Reference Begg and Mazumdar 34 , Reference Egger, Smith and Schneider 35 ). We excluded one small study( Reference Pollack, Buck Louis and Sundaram 36 ) with implausibly small se for effect estimates in the assessment of publication bias. We further conducted trim-and-fill analysis to generate summary effect estimates adjusted for publication bias( Reference Duval and Tweedie 37 ). All tests were performed using the statistical software package STATA version 11·2 and two-sided P values <0·05 were considered statistically significant.
The flow diagram with details of the study selection is shown in Fig. 1. We included fourteen prospective studies( Reference Bech 12 , Reference Klebanoff, Levine and DerSimonian 23 , Reference Wisborg, Kesmodel and Bech 26 , Reference Pollack, Buck Louis and Sundaram 36 , Reference Srisuphan and Bracken 38 – Reference Greenwood, Alwan and Boylan 47 ) on caffeine intake and pregnancy loss involving 130 456 participants and 3429 cases of pregnancy loss (Table 1). One of the studies( Reference Srisuphan and Bracken 38 ) was included for only qualitative review because it did not report data for more than two categories of caffeine intake. All studies in the present meta-analysis were conducted in the USA or Europe.
Figure 2 shows the RR for the association between maternal caffeine intake and pregnancy loss. The summary RR was 1·72 (95 % CI 1·40, 2·13) for very high caffeine intake (≥700 mg/d), 1·40 (95 % CI 1·16, 1·68) for high caffeine intake (350–699 mg/d), 1·16 (95 % CI 0·94, 1·41) for moderate caffeine intake (150–349 mg/d) and 1·02 (95 % CI 0·85, 1·24) for low caffeine intake (50–149 mg/d), as compared with the reference category with no or very low caffeine intake. The heterogeneity in study results was low to moderate: the I 2 was 28·3 % for low caffeine intake, 49·6 % for moderate caffeine intake, 18·6 % for high caffeine intake and 0 % for very high caffeine intake. Srisuphan and Bracken( Reference Srisuphan and Bracken 38 ) also reported a significant association between caffeine intake and risk of late spontaneous abortion, but their study could not be included in the meta-analysis because it considered only two categories of intake.
We assumed a linear relationship in the dose–response meta-analysis between maternal caffeine intake and pregnancy loss, because there was no evidence of statistically significant departure from linearity (P=0·44). The summary RR was 1·07 (95 % CI 1·03, 1·12; Table 2 and Fig. 3) per 100 mg/d. To facilitate comparison with the categorical analysis this can also be expressed as 1·23 (95 % CI 1·09, 1·39) per 300 mg/d or 1·42 (95 % CI 1·16, 1·74) per 500 mg/d increment of maternal caffeine intake. The heterogeneity in study results was higher (I 2=80·9 %) for this analysis than for the comparisons of specific categories of caffeine consumption.
RR, relative risk; NA, not available; Ref., reference.
* P value was obtained by modelling year of publication and median of assigned doses as continuous variables.
† Mean age <30 years or ≥30 years. If mean age is not available, classification was based on whether majority of the population (>50 %) is <30 years or ≥30 years.
‡ Total number of studies is more than thirteen because some studies reported additional (usable) results for a different exposure period.
§ Fine adjustment for smoking refers to studies that adjusted for amount of smoking or studies that adjusted for smoking using a biomarker; crude adjustment refers to studies that did not adjust for amount of smoking; ‘Not applicable’ refers to studies that presented estimates without adjustment for smoking as adjustment for smoking did not change the results substantially.
The summary RR did not differ substantially in various subgroups (Table 2), except for mean baseline age of the participants. The association was stronger (P for interaction=0·02) in studies of older women (mean age ≥30 years, summary RR=1·23, 95 % CI 1·09, 1·38) than in studies of younger women (mean age <30 years, summary RR=1·05, 95 % CI 1·04, 1·07; Table 2). In addition, there was a trend of borderline significance (P for interaction=0·05) towards a stronger effect of maternal caffeine intake assessed during pregnancy (summary RR=1·11, 95 % CI 1·05, 1·17) as compared with caffeine intake assessed during the pre-pregnancy period (summary RR=1·02, 95 % CI 0·97, 1·07). However, when we modelled both study characteristics in the meta-regression analysis, only maternal age ≥30 years remained significant (P=0·02), while the interaction with period of assessment was attenuated (P=0·52).
In a sensitivity analysis, the summary RR ranged from 1·06 (95 % CI 1·02, 1·10) to 1·08 (95 % CI 1·03, 1·14) per 100 mg/d increment in maternal caffeine intake when we omitted studies one at a time. When we excluded studies focusing on stillbirth or a combination of miscarriage and stillbirth, the pooled effect estimates were not materially different in both the dose–response (summary RR=1·08, 95 % CI 1·04, 1·13) and category-based meta-analysis, and the conclusion remained the same. There was a suggestion of small-study effect such as publication bias based on visual inspection of the funnel plot (Fig. 4) and in both Egger’s test (P<0·01) and Begg’s test (P=0·047). However, the summary effect estimate remained significant in the subgroup of the larger studies (RR=1·06 per 100 mg/d, 95 % CI 1·03, 1·10; Table 2) or when we conducted the trim-and-fill analysis (RR=1·06, 95 % CI 1·02, 1·10) that imputed six inferred missing studies.
The findings from the current meta-analysis of prospective studies suggest that maternal caffeine intake is associated with a higher risk of pregnancy loss. High maternal intake of caffeine (350–699 mg/d) was associated with a 40 % higher risk of pregnancy loss. In our dose–response analysis, a 100 mg/d increment in maternal caffeine intake (~1 cup of coffee) was associated with a 7 % higher risk for pregnancy loss. However, our results should be interpreted with caution in consideration of methodological limitations of the original studies and other potential issues such as residual confounding and publication bias that may have influenced our results.
Most previous reviews on maternal caffeine intake and pregnancy loss included only a qualitative summary of the evidence( Reference Peck, Leviton and Cowan 17 , Reference Matijasevich, Santos and Barros 48 , Reference Signorello and McLaughlin 49 ), used estimates that were not adjusted for potential confounders( Reference Fernandes, Sabharwal and Smiley 50 ) or included only five studies on preconception caffeine intake( Reference Lassi, Imam and Dean 51 ). A recent meta-analysis reported significant associations stronger than those observed in our meta-analysis between higher maternal caffeine intakes and higher risks of miscarriage (summary RR=1·14 per 100 mg caffeine/d, 95 % CI 1·10, 1·19 in that study compared with 1·08, 95 % CI 1·04, 1·13 in our study) and stillbirth (summary RR=1·19, 95 % CI 1·05, 1·35 in that study compared with 1·09, 95 % CI 1·02, 1·16 in our study)( Reference Greenwood, Thatcher and Ye 18 ). However, that meta-analysis also included retrospective studies which are more prone to recall and selection bias. Similar to our study, high heterogeneity was reported in the dose–response analysis and publication bias may have affected the observed associations. We also conducted a complementary analysis using specific categories of caffeine consumption which resulted in substantially lower heterogeneity. This suggests that the estimation of dose–response relationships based on the results for categories of caffeine consumption reported in the original studies may have introduced additional heterogeneity. The correlation matrices of the adjusted and unadjusted log RR are assumed to be approximately equal in the GLST method proposed by Greenland and Longnecker, and approximation using this method tends to be unstable if there is small number of cases in the referent exposure level( Reference Orsini, Li and Wolk 52 ). To date, there has been no randomized controlled trial conducted to evaluate the effect of a reduction in maternal caffeine intake on pregnancy loss( Reference Jahanfar and Jaafar 53 , Reference Jahanfar and Sharifah 54 ).
Our meta-analysis has several strengths. We searched databases with no language restriction to increase the completeness of the identification of studies. Moreover, we included only prospective studies and thus reduced the influence of selection bias, recall bias and reverse causation. The included studies were reasonably large, with study size of more than 1000 participants in ten out of fourteen studies. Lastly, we did not only compare the highest categories with the lowest categories but divided maternal caffeine intake into five levels and also conducted a dose–response analysis.
Our study results should be interpreted with consideration of several potential limitations. First, our results could have been affected by measurement error of caffeine intake assessment in the original studies. The included studies relied mostly on questionnaires for caffeine assessment. However, validation studies have shown that self-reports of major sources of caffeine such as coffee and tea are in general reasonably accurate and reliable( Reference James, Bruce and Lader 55 , Reference Grosso, Triche and Benowitz 56 ). Another potential limitation is that two of the included studies conducted in Denmark used coffee consumption as the proxy of total caffeine intake. However, coffee consumption is high in Denmark and in a similar Swedish study coffee was the predominant source (accounting for 76 %) of all caffeine ingested by pregnant women( Reference Cnattingius, Signorello and Annerén 57 ). In our stratified analyses, the summary estimates did not differ substantially by dietary assessment method or by assessment of coffee v. total caffeine intake. The caffeine assessments in the remaining studies were reasonably comprehensive. Most studies considered at least the intakes of coffee, tea and caffeinated soft drinks( Reference Pollack, Buck Louis and Sundaram 36 , Reference Srisuphan and Bracken 38 – Reference Wen, Shu and Jacobs 43 , Reference Savitz, Chan and Herring 45 – Reference Greenwood, Alwan and Boylan 47 ), while some also additionally considered other sources such as cocoa products( Reference Mills, Holmes and Aarons 40 , Reference Wen, Shu and Jacobs 43 , Reference Weng, Odouli and Li 46 , Reference Greenwood, Alwan and Boylan 47 ) and caffeine-containing medication( Reference Srisuphan and Bracken 38 , Reference Mills, Holmes and Aarons 40 , Reference Greenwood, Alwan and Boylan 47 ). Energy drinks containing relatively high amounts of caffeine have increased in popularity, especially since the mid-2000s( Reference Heckman, Sherry and De Mejia 58 , Reference Oddy and O’Sullivan 59 ). However, most studies included in our meta-analysis were conducted before that period. Furthermore, the contribution of energy drinks to total caffeine intakes was found to be small in studies conducted in pregnant women or women of childbearing age( Reference Derbyshire and Abdula 60 , Reference Knight, Knight and Mitchell 61 ).
Another potential limitation of our meta-analysis is residual confounding resulting from unmeasured or imperfectly measured covariates. A majority of the included studies adjusted for potential confounders rather comprehensively. We only included studies that considered potential confounding by smoking, either by adjusting for it in multivariable models or reporting that such adjustment had minimal influence on the results. Moreover, most studies adjusted for maternal age, parity, alcohol intake and socio-economic status.
Smoking tends to correlate with caffeine intake( Reference Sengpiel, Elind and Bacelis 9 , Reference Mills, Holmes and Aarons 40 ) and is a known risk factor for pregnancy loss( Reference Cnattingius 62 , Reference Pineles, Park and Samet 63 ). Smoking is therefore a potentially important confounder of the association between caffeine intake and pregnancy loss. The six included studies( Reference Bech 12 , Reference Wisborg, Kesmodel and Bech 26 , Reference Pollack, Buck Louis and Sundaram 36 , Reference Mills, Holmes and Aarons 40 , Reference Tolstrup 44 , Reference Weng, Odouli and Li 46 ) that investigated a possible interaction between caffeine intake and smoking in relation to pregnancy loss either did not suggest differences in association between smokers and non-smokers( Reference Bech 12 , Reference Wisborg, Kesmodel and Bech 26 , Reference Pollack, Buck Louis and Sundaram 36 , Reference Mills, Holmes and Aarons 40 , Reference Tolstrup 44 ) or suggested a stronger association in non-smokers( Reference Weng, Odouli and Li 46 ). It is reassuring that the association between caffeine intake and pregnancy loss was observed in non-smokers, because residual confounding by intensity of smoking is not a concern in this group. However, non-disclosure rate of active smoking was higher among pregnant women compared with non-pregnant women in a study comparing self-reported and biomarker-determined smoking status( Reference Dietz, Homa and England 64 ). Thus, residual confounding by smoking may still be a concern due to under-reporting of smoking in pregnant women. Nevertheless, in Greenwood et al.’s study higher maternal caffeine intake was associated with a higher risk of late miscarriage and stillbirth even after adjustment for salivary cotinine, a biomarker for smoking( Reference Greenwood, Alwan and Boylan 47 ).
Pregnancy symptoms including nausea, vomiting, and aversions to smells and taste are more common in healthy pregnancies than in pregnancies that end in spontaneous abortion( Reference Peck, Leviton and Cowan 17 ). Women with viable pregnancies may be more likely to decrease their caffeine consumption in response to pregnancy symptoms( Reference Peck, Leviton and Cowan 17 ). Pregnancy symptoms can affect the interpretation of the relationship between caffeine and pregnancy loss in two ways. First, pregnancy symptoms are associated with both pregnancy loss and maternal caffeine intake and thus potentially are a confounder. Second, loss of fetal viability (which precedes the actual detection of pregnancy loss) can result in reduced pregnancy symptoms and subsequently increased caffeine intake; this scenario represents reverse causality. If pregnancy symptoms confound the relationship between caffeine and pregnancy loss, their influence can be studied through adjustment or stratification for these symptoms. Wen et al. reported a substantial association between caffeine intake measured after nausea occurrence and the risk of spontaneous abortion in women with nausea, but not for caffeine intake assessed earlier or in women without nausea( Reference Wen, Shu and Jacobs 43 ). In contrast, Klebanoff et al. observed similar associations in women who did or did not experience vomiting( Reference Klebanoff, Levine and DerSimonian 23 ). Other studies that investigated the influence of pregnancy signals using self-reported information on nausea, vomiting and fatigue did not support the argument that confounding by pregnancy-related symptoms results in spurious associations between maternal caffeine intake and the risk of spontaneous abortion( Reference Savitz, Chan and Herring 45 , Reference Weng, Odouli and Li 46 , Reference Cnattingius, Signorello and Annerén 57 ). Furthermore, the association between maternal caffeine intake and pregnancy loss was not significantly different for studies that did and did not adjust for nausea and vomiting in our meta-regression analysis (Table 2). Nevertheless, assessment of relevant pregnancy symptoms is unlikely to be without measurement error and residual confounding by symptoms reflecting fetal viability thus cannot be completely excluded. Moreover, the possibility that higher caffeine intake is a consequence rather than a cause of pregnancy loss (reverse causality) cannot be excluded. A study with detailed serial measurement of both caffeine intake and fetal viability based on ultrasounds may help to disentangle the cause-and-effect sequence, but such measurements are unlikely to be feasible in big epidemiological studies.
In addition to methodological limitations of the original studies, our results may have been affected by small-study effect such as publication bias as suggested by the funnel plot and statistical tests. Publication bias is only one of the possible reasons for small-study effect and there can be a real difference in effect size in smaller studies due to methodological or clinical diversity. It is possible that smaller studies have a larger effect size due to more complete and comprehensive measurement of caffeine intake. Nevertheless, in our stratified analysis the association between caffeine intake and pregnancy loss was similar in the larger studies that are less likely to be affected by this bias, but as in any meta-analysis the possibility of the bias cannot be completely excluded even for the subset of larger studies. If the observed small-study effect is really due to publication bias, our trim-and-fill analysis suggested that the bias did not affect the pooled effect estimate substantially.
The exact mechanism through which caffeine may increase risk of pregnancy loss remains unclear. Caffeine can increase the release of catecholamines, which may lead to vasoconstriction in the uteroplacental circulation and fetal hypoxia and subsequently affect fetal development and growth( Reference Bech 12 , Reference Kirkinen, Jouppila and Koivula 13 ). Indeed, after maternal ingestion of just 200 mg of caffeine, a 25 % reduction in intervillous placental blood flow has been documented( 6 , Reference Kirkinen, Jouppila and Koivula 13 ). Another proposed mechanism is that caffeine increases the cellular concentration of cyclic adenosine monophosphate (cAMP) by inhibiting an enzyme (phosphodiesterase) responsible for the breakdown of cAMP( Reference Sengpiel, Elind and Bacelis 9 , Reference Weathersbee and Lodge 65 ). A build-up of cAMP may affect fetal growth through its influence on cell division or catecholamine-mediated vasoconstriction( Reference Grosso, Rosenberg and Belanger 66 , Reference Soyka 67 ). Furthermore, caffeine may also affect the fetal cardiovascular system directly by increasing fetal heart rate accelerations and causing tachycardia( Reference Wisborg, Kesmodel and Bech 26 , Reference Buscicchio, Piemontese and Gentilucci 68 , Reference Resch, Papp and Gyöngyösi 69 ). As mentioned before, spontaneous abortion and stillbirth may have different aetiology. Caffeine may influence these outcomes through a common mechanism such as vascular disorder or via different pathways, but there is limited evidence to postulate further. If maternal caffeine intake is indeed causally associated with pregnancy loss, the effect can also be possibly exerted by its metabolites. These may have important implications as wide inter-individual variation in caffeine metabolism due to either genetic variation (e.g. polymorphism in CYP1A2) or lifestyle factors (e.g. smoking) has been reported( Reference Grosso and Bracken 70 ). Thus, in order to provide a more complete picture on caffeine metabolism, future studies should consider assaying for circulating caffeine, its metabolites and genotype information on top of well-planned caffeine intake and lifestyle factors measurements.
It is not clear why the association between caffeine intake and pregnancy loss was more pronounced in studies of older women than in studies of younger women and this observation requires confirmation in further studies with individual-level data on age. It may be that older mothers drink more coffee( Reference Weng, Odouli and Li 46 , Reference Olsen, Overvad and Frische 71 ) or that older age inherently exacerbates the negative influence of caffeine on pathways leading to pregnancy loss. It is known that advancing maternal age increases risk of fetal chromosomal abnormalities and thus results in higher risk of spontaneous abortion. However, caffeine is generally not considered to be mutagenic and cytotoxic at usual human exposure level( Reference Brent, Christian and Diener 72 ).
The present systematic review and meta-analysis supports the hypothesis that high maternal caffeine intake during pregnancy is associated with a higher risk of pregnancy loss. The results of recent meta-analyses also indicate that higher maternal coffee consumption may increase the risk of low birth weight in offspring( Reference Greenwood, Thatcher and Ye 18 , Reference Chen, Wu and Neelakantan 73 ). Currently, the WHO recommends a caffeine intake of less than 300 mg (~3 cups of coffee) per day. Our results suggested that high caffeine intake (~3·5 to 7 cups of coffee) and very high caffeine intake (≥7 cups of coffee) were associated with a substantially higher risk of pregnancy loss. Our results may have been affected by publication bias and the possibility of residual confounding by smoking or pregnancy symptoms cannot be ruled out. Nevertheless, in combination with the biological plausibility of adverse effects on the fetus and evidence for effects of maternal caffeine intake on reduced fetal growth, it appears prudent to adhere to public recommendations to limit caffeine intake during pregnancy to low levels.
Financial support: The project was supported by Saw Swee Hock School of Public Health, National University of Singapore, Singapore. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Conflict of interest: None. Authorship: L.-W.C., M.F.-F.C. and R.M.v.D. conceived the study. L.-W.C. and Y.W. searched the databases and screened studies according to inclusion and exclusion criteria. L.-W.C., Y.W., M.F.-F.C. and R.M.v.D. developed the search strategy. A.P. and N.N. gave advice on meta-analysis methodology. L.-W.C. analysed the data. A.P. and R.M.v.D. helped in the interpretation of the results. L.-W.C. wrote the draft of the paper. L.-W.C., Y.W., N.N., M.F.-F.C., A.P. and R.M.v.D. were involved in critical reviewing and revision of the manuscript. All authors approved the final version of the manuscript. Ethics of human subject participation: Not applicable.
To view supplementary material for this article, please visit http://dx.doi.org/10.1017/S1368980015002463